A boosted classifier for integrating multiple fields of view: Breast cancer grading in histopathology

The ability to accurately interpret large image scenes is often dependent on the ability to extract relevant contextual, domain-specific information from different parts of the scene. Traditionally, techniques such as multi-scale (i.e. multi-resolution) frameworks and hierarchical classifiers have been used to analyze large images. In this paper we present a novel framework that classifies entire images based on quantitative features extracted from fields of view (FOVs) of varying sizes (i.e. multi-FOV scheme). The boosted multi-FOV classifier is subsequently applied to the task of computerized breast cancer grading (low vs. high) in digitized, whole-slide histopathology images. First an image is split up into many FOVs at different FOV sizes. In each FOV, cancer nuclei are automatically detected and used to construct graphs (Voronoi Diagram, Delaunay Triangulation, Minimum Spanning Tree). Features describing spatial arrangement of the nuclei are extracted and used to train a boosted classifier that predicts image class for each FOV size. The resulting predictions are then passed to the boosted multi-FOV classifier, which weights individual FOV sizes based on their ability to discriminate low and high grade BCa. Using slides from 55 patients, boosted classifiers were constructed using both multi-FOV and multi-scale frameworks, resulting in area under the receiver operating characteristic curve (AUC) values of 0.816 and 0.791, respectively.